Multi-object Detection In Complex Traffic Scenes Based On Deep Learning

Traffic scene object detection has always been a focused and difficult problem in the fields of automatic driving and intelligent transportation.Because the object detection of traffic scenes is very complicated,it is easily affected by weather conditions,light intensity,truncation and occlusion,and faces enormous challenges in the feature extraction process.In recent years,with the development of deep learning technology,especially the powerful feature extraction ability of convolutional neural network,a large number of object detection algorithms based on deep learning have emerged.The object detection algorithm for deep learning is more rigorous than traditional algorithms.Bringing new opportunities for object detection in traffic scenes.This thesis studies the problems of dense objects,mutual occlusion phenomenon and missing detection caused by too many small objects in traffic scenes.The main work is as follows:(1)Analysing the advantages of deep learning in traffic scene object detection,comparing two kinds of deep learning object detection algorithms based on region proposal and bounding box regression,considering the accuracy and real-time of the algorithm,choosing YOLOv3 algorithm as the basic model.(2)Designing a dense connection algorithm dense?YOLOv3,based on Dense Net network,which is aimed at the peoblems of low parameter utilization of the residual network Res Net in the YOLOv3 object detection algorithm.Compared with Res Net,Dense Net adds feature reuse.The feature parameters are effectively utilized to avoid over-fitting problems caused by the network being too deep.Embedding a attention mechanism SENet in the Densenet network structure,SENet can effectively prevent the interference of background information and learn the importance of each channel feature of the feature map.Finally,the data set of the traffic scene is re-clustered to obtain the optimal anchor suitable for the paper.The network model is retrained and tested by transfer learning.The improved network SE?dense YOLOv3 can identify the location of object more accurately and eliminate interference,which has strong robustness.The average accuracy rate on the INRIA pedestrian test data set and the KITTI data set increased by 5.7% and 6.32%.(3)Aiming at the intensive object in complex traffic scenes,the missed object caused by the proximity of the object,and the multiple objects in the same detection box,a objected occlusion regression loss function Occlusion Loss is proposed on SE?dense YOLOv3.Occlusion Loss has two functions: one is to guide the neural network learning detection box and the groundtruth matching degree to obtain more accurate position information;the second is to reduce the number of detected objects in one detection box as much as possible after learning the position information..The training and testing were carried out on the re-divided traffic scene dataset KITTI.The experimental results show that the improved YOLOv3 is more ac-curate in positioning,and the average accuracy is 3.29% higher than the original algorithm.